Functions for adult neurogenesis in memory: An introduction to the neurocomputational approach and to its contribution

Abstract

Until recently, it was believed that the introduction of new neurons in neuronal networks was incompatible with memory function. Since the rediscovery of adult hippocampal neurogenesis, behavioral data demonstrate that adult neurogenesis is required for memory processing. We examine neurocomputational studies to identify which basic mechanisms involved in memory might be mediated by adult neurogenesis. Mainly, adult neurogenesis might be involved in the reduction of catastrophic interference and in a time-related pattern separation function. Artificial neuronal networks suggest that the selective recruitment of new-born or old neurons is not stochastic, but depends on environmental requirements. This leads us to propose the novel concept of “soft-supervision”. Soft-supervision would be a biologically plausible process, by which the environment is able to influence activation and learning rules of neurons differentially.

Introduction

Memory processes include information acquisition, encoding, storing, retention, and retrieval. Given that memory is never static in living animals, this also includes information processing, which allows acquired information to be used to infer new information and use it in various situations. Because memory implies information storage, it has been hypothesized that information is stored by long-term synaptic changes, and is “represented” in the brain by the activity of specific cells or sets of cells. In other words, the informational properties of – hypothetically – coding cells result from the modification of synaptic strengths among cells of the network. In a preliminary analysis, the challenge raised by the discovery of adult neurogenesis in structures which are known to be involved in memory, i.e. the dentate gyrus and the olfactory bulb, is that the introduction of new neurons into a memory network may weaken its stability because a “stable population of neurons may be a biological necessity in an organism whose survival relies on learned behavior acquired over a long period of time” [1], [2], [3].

In an evolutionist perspective, one should ask why adult neurogenesis is present in the dentate gyrus (DG) and olfactory bulb and has not been found in other areas: is adult neurogenesis the non-functional remains of a decreasing phenomenon which was present in the whole brain? Is adult neurogenesis useful in these two structures (or in a minority of structures), and not in the others so that it is specifically maintained in these specific locations and not in the others? And if yes, what is this use? Finally, is neurogenesis a “new” phenomenon which appeared in these structures first, but which may be useful in others, and spread in the brain with the future stages of evolution?

The basic aim of the neurocomputational approach is to identify how brain structures can generate specific functions, thanks to their input, output, internal architecture and rules governing neuronal interaction. The corollary is that it allows to test whether the hypothetical functions of a brain structure are compatible with the internal architecture of the structure.

Interestingly, the concept of function that a structure can produce has been modified by the neurocomputational approach. Without it, one can suppose that brain structures produce integrated psychological functions. For example, the hippocampus has been hypothesized to be responsible for integrated cognitive functions such as cognitive mapping [4] or working memory [5]. These theories are often in opposition. On the contrary, the neurocomputational view emphasizes that each structure, or even each part of a structure performs several basic computational functions (or “rudimentary functions” [6]), which may be used or necessary for various integrated cognitive functions. The activation of networks of structures, i.e. networks of computational functions, would produce the integrated psychological function. Examples of computational functions may be the hetero-association which consists in associating two different stimuli [7], learning sequences of stimuli (when the first stimulus is presented, the other stimuli of the sequence are recalled [7]), pattern classification (also called “discriminant function” [8]), pattern completion (the ability to retrieve the whole code of a stimulus when a subpart of it is presented [7]), or pattern separation (also called “orthogonalization of codes”; when two stimuli are represented by two highly similar sets of neurons, pattern separation consists in replacing their neuronal representation by two other non-overlapping or less-overlapping sets of neurons [9]).

Within the concept of rudimentary functions, the neurocomputational approach may help to identify the usefulness and/or the negative consequences of the adult neurogenesis phenomenon. Schematically, neurocomputational theories suggest that adult neurogenesis may interfere with memory at two levels: it may interfere with neural representations of information and/or with the mechanisms which make it possible to learn and process the representations. In neurocomputational models, information is implemented through an input pathway and is processed by the network thanks to a specific architecture and intercellular connection rules. The result of the computation can be observed in a set of “potentially coding cells”, within which each cell can be differentially activated. The specific pattern of activation of the set of potentially coding cells forms a code, representing computed information. This code is a symbolic representation of information. Schematically, an observer who is aware of the information stored by the network and how it is coded, is able to identify which memory representation is presently active in the network, simply by observing the pattern of activity of the set of potentially coding cells. This code may be used for further processing stages. It may be the activity of a specific cell (unicellular coding), or of specific sets of cells (multicellular coding), and some other types of codes have also been hypothesized, such as spatial patterns of activity [8] that may be recorded in the electro-encephalogram, as exemplified by the early hypotheses of Freeman and Skarda [10].

Parts of the brain where recording of some kinds of activity, most often cellular activity, makes it possible to identify stored information from the environment, may reveal the presence of stored codes. For example, recording of place cells in the hippocampus, “hand” and “face” cells in the inferior temporal area [11], cells specifically activated by colors [12], [13], speed [14], or grid cells [15] may reveal coding activities. In that view, stored codes remain unchanged for a significant time, representing information about the world, and are used to process information.

In animals, memory is not a static phenomenon. Known information can be transformed or used to infer new information. Computational theories emphasize that neural networks perform these operations by processing neural codes representing information. Processing may have a bearing on the representation itself, e.g. in order to complete a truncated code (pattern completion), to detect familiarity (to identify information as known or novel). Processing may also use these representations in order to infer new information, e.g. representations of local places may be used to create spatial maps, then spatial maps may be used to infer short-cuts. In some artificial networks (multilayer networks where output cells are not interconnected; cf. Fig. 1A), coding and processing areas are well separated. Potentially coding cells are located in the output layer and processing is performed in “hidden layers” and connections. In other networks, the distinction between coding and processing is not so clear, as in Hopfield's networks [16] where the same layer has both a coding and processing role, given the connections between units (cf. Fig. 1B).

In the brain, it is likely that representation and coding functions are mixed in most structures. Particularly, place fields have been found in granule cells of the DG [17], revealing the presence of a coding level, and DG is also hypothesized to perform some processing such as pattern separation [9]. Also, given the difference between receptive fields found in the V1 area of the visual cortex (edge, motion, size detectors) and the high specificity of response of the cells in the temporal inferior area (face cells, hand cells), it is likely that the intermediate areas in this pathway (namely V2 and V4) have a processing function in the hierarchical processing of visual information. Electrophysiological studies have permitted to identify some properties of stimuli which are able to activate V2 and V4 cells [18], [19], [20]. This suggests that codes may exist within these structures together with a processing function.

As demonstrated by the development of hierarchical networks, which can be defined as network of several artificial networks, interlacing between coding populations and processing operations appears adaptive [21], [22].

In order to analyze the consequences of adult neurogenesis in cognition, we will successively analyze its consequences when new neurons are introduced in the potentially coding set of cells, and then when they are introduced in neural populations where processing happens.